The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python's subprocess module to execute the claude CLI and local evaluation scripts. This is a core functional requirement for verifying how other skills are triggered and performed.
[EXTERNAL_DOWNLOADS]: The eval-viewer/viewer.html file references external assets including Google Fonts and the SheetJS library (cdn.sheetjs.com) for rendering spreadsheets in the report. These are well-known services used for UI presentation.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it processes untrusted user prompts and subagent outputs during the skill improvement loop. It uses structured XML-like delimiters (e.g., <skill_content>, <attempt>) to separate instructions from data during the description optimization process.
Ingestion points: evals/evals.json, feedback.json, and subagent transcripts.
Boundary markers: Employs explicit XML-style tags and imperative instructions to Claude to focus on intent generalization.
Capability inventory: Includes file system access (read/write), subprocess execution, and local HTTP server hosting.
Sanitization: Employs HTML escaping in report generation and validates SKILL.md frontmatter formatting.

skill-creator